Experiment configuration

class previsionio.experiment_config.ColumnConfig(target_column: Optional[str] = None, filename_column: Optional[str] = None, id_column: Optional[str] = None, fold_column: Optional[str] = None, weight_column: Optional[str] = None, time_column: Optional[str] = None, group_columns: Optional[List[str]] = None, apriori_columns: Optional[List[str]] = None, drop_list: Optional[List[str]] = None)

Column configuration for starting an experiment: this object defines the role of specific columns in the dataset (and optionally the list of columns to drop).

  • target_column (str, optional) – Name of the target column in the dataset
  • id_column (str, optional) – Name of the id column in the dataset that does not have any signal and will be ignored for computation
  • fold_column (str, optional) – Name of the fold column used that should be used to compute the various folds in the dataset
  • weight_column (str, optional) – Name of the weight column used to assign non-equal importance weights to the various rows in the dataset
  • filename_column (str, optional) – Name of the filename column in the dataset for an image-based experiment
  • time_column (str, optional) – Name of the time column in the dataset for a timeseries experiment
  • group_columns (list(str), optional) – Names of the columns in the dataset that define a unique time serie for a timeseries experiment
  • apriori_columns (list(str), optional) – Names of the columns that are known a priori in the dataset for a timeseries experiment
  • drop_list (list(str), optional) – Names of all the columns that should be dropped from the dataset while training the experiment
class previsionio.experiment_config.ExperimentState

Possible state of an Experiment in Prevision.io

Done = 'done'

The experiment finished properly

Failed = 'failed'

The experiment finished with an error

Pending = 'pending'

The experiment is waiting for hardware ressources

Running = 'running'

The experiment is still running

class previsionio.experiment_config.Provider

Type of provider of model described in Prevision.io

External = 'external'

Model provided by user

Prevision = 'prevision-auto-ml'

Model created by Prevision.io

class previsionio.experiment_config.TrainingConfig(profile: previsionio.experiment_config.Profile = <Profile.Quick: 'quick'>, advanced_models: List[previsionio.experiment_config.AdvancedModel] = [<AdvancedModel.XGBoost: 'XGB'>, <AdvancedModel.LinReg: 'LR'>], normal_models: List[previsionio.experiment_config.NormalModel] = [<NormalModel.XGBoost: 'XGB'>, <NormalModel.LinReg: 'LR'>], simple_models: List[previsionio.experiment_config.SimpleModel] = [], features: List[previsionio.experiment_config.Feature] = [<Feature.Frequency: 'freq'>, <Feature.TargetEncoding: 'tenc'>, <Feature.Counts: 'Counter'>], with_blend: bool = False, feature_time_seconds: int = 3600, feature_number_kept: Optional[int] = None)

Training configuration that holds the relevant data for an experiment description: the wanted feature engineering, the selected models, the training speed…

  • profile (Profile) –

    Type of training profile to use:

    • ”quick”: this profile runs very fast but has a lower performance (it is recommended for early trials)
    • ”advanced”: this profile runs slower but has increased performance (it is usually for optimization steps at the end of your project)
    • the “normal” profile is something in-between to help you investigate an interesting result
  • advanced_models (list(AdvancedModel), optional) – Names of the advanced models to use in the experiment (among: “LR”, “RF”, “ET”, “XGB”, “LGB”, “CB” and “NN”). The advanced models will be hyperparametrized, resulting in a more accurate modelization at the cost of a longer training time.
  • normal_models (list(NormalModel), optional) – Names of the (normal) models to use in the experiment (among: “LR”, “RF”, “ET”, “XGB”, “LGB”, “CB”, ‘NB’ and “NN”). The normal models only use default parameters.
  • simple_models (list(SimpleModel), optional) – Names of the (simple) models to use in the experiment (among: “LR” and “DT”). These models are easy to ineterpret and fast to train but only offer a limited modelization complexity.
  • features (list(Feature), optional) – Names of the feature engineering modules to use (among: “Counter”, “Date”, “freq”, “text_tfidf”, “text_word2vec”, “text_embedding”, “tenc”, “ee”, “poly”, “pca” and “kmean”)
  • with_blend (bool, optional) – If true, Prevision.io’s pipeline will add “blend” models at the end of the training by cherry-picking already trained models and fine-tuning hyperparameters (usually gives even better performance)
  • feature_time_seconds (int, optional) – feature selection take at most fsel_time in seconds
  • feature_number_kept (int, optional) – a feature selection algorithm is launched to keep at most feature_number_kept features
class previsionio.experiment_config.TypeProblem

Type of supervised problems available with Prevision.io.

Classification = 'classification'

Prediction using classification approach, for when the output variable is a category

MultiClassification = 'multiclassification'

Prediction using classification approach, for when the output variable many categories

ObjectDetection = 'object-detection'

Detection of pattern in images

Regression = 'regression'

Prediction using regression problem, for when the output variable is a real or continuous value

TextSimilarity = 'text-similarity'

Ranking of texts by keywords

class previsionio.experiment_config.YesOrNo

An enumeration.

class previsionio.experiment_config.YesOrNoOrAuto

An enumeration.